Research map / Federated learning and privacy / Differential privacy
On Differentially Private Mechanisms for Linear Regression
Bardiya Aryanfard, Monika Henzinger, Farhood Rostamkhani
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Studies differentially private least-squares linear regression in the continual observation setting, where estimates of optimal parameters are output at every time step of a data stream.
Problem
Differentially private linear regression under continual observation, where the mechanism must output an estimate of the optimal least-squares parameters at each time step of a stream of n data points (x_i, y_i).
Approach
Abstract truncated: the proposed mechanism for private least squares in the continual observation setting is not described in the provided text, so the core idea cannot be stated from the abstract alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| scale the adaptive regularization by c=1/16 instead of c=1 (ablation) | ContinualAdaSSP · Year, Household, ACS Income (eps 1 and 2) | average loss: 6.7 -> 2.7 (Year, eps=1) | Improved |
| use adaptive regularization that decreases over time instead of PQFTL's regularization growing with t | ContinualAdaSSP · Year, ACS Income | average loss: 7.1 -> 6.7 (Year, eps=1); 2.7 -> 2.5 (ACS, eps=1) | Improved |
| rerun the batch DP regression at every time step (with composition) instead of continual mechanism | Batch AdaSSP · Year, Household, ACS Income | average loss: 7.1 -> 1.9 x 10^3 vs PQFTL (Year, eps=1) | Hurt |
| increase privacy budget epsilon from 1 to 2 (ablation) | ContinualAdaSSP · Year | average loss: 2.7 -> 1.8 (c=1/16) | Improved |
| increase privacy budget epsilon from 1 to 2 (ablation) | PQFTL · Year, Household, ACS Income | average loss: 7.1 -> 7.1 (Year) | No clear effect |
Closest papers by meaning
- Private Gradient Descent for Linear Regression: Tighter Error Bounds and Instance-Specific Uncertainty Estimation (ICML 2024 · similarity 0.89)
- Truthful High Dimensional Sparse Linear Regression (NeurIPS 2024 · similarity 0.88)
- Improved Analysis of Sparse Linear Regression in Local Differential Privacy Model (ICLR 2024 · similarity 0.87)
- Better Locally Private Sparse Estimation Given Multiple Samples Per User (ICML 2024 · similarity 0.87)
- Revisiting Differentially Private ReLU Regression (NeurIPS 2024 · similarity 0.87)
- Instance-Specific Asymmetric Sensitivity in Differential Privacy (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
